INTERNAL fMRI experiments page

This page is for internal use by members of this project.

It can also be accessed from "http://www.cs.cmu.edu/afs/cs/project/theo-35/fmri/www/" inside the ".cmu.edu" domain.

NEW!: Summary of filetypes and data manipulation utilities and instructions on how to use them  (Not finished yet!)

Data:

Data sets:

• Data from right and left broca and temporal regions from Just/Carpenter language experiments. See data on /afs/cs/project/theo-35/fmri/ and the associated /afs/cs/project/theo-35/fmri/readme file.
• Which voxels are relevant according to Just/Carpenter
• List of sentences and questions presented to subjects
• t test info - for subject 02882 for subject 02930

File naming conventions and file layouts:

• for raw data files and intermediate examples files
• bad and other image info layout for intermediate example files

Stimulus and sentence info:

• Description
• Sentences without questions grouped by condition - html, text, original ms word file
• Full stimulus file - cleaned up text, original file

Code:

• Description of data manipulation utilities
• pff2ppm
• Voxel cropper
• Voxel plotter. And a sample input file. Results of plotting neural network NRMSE: full brain and close up and full brain inverse and close up inverse .

Experimental results:

• visual displays of means and variances - for subject 02882: mean, var, stdev - for subject 02930: mean, var, stdev
• Using neural network to predict left broca(t+1) from left broca(t).Initial results show the average RMS error for neural net is 35.1, compared to 47.3 when using voxel mean as the prediction, and 44.7 when using the previous voxel value as the prediction.

More detailed results on a later run include the partial derivatives of the predictions.
• Using linear regression to predict left broca(t) from left broca(t-1,..,t-N). N varies from 1 to 10.Results show that the RMS error on the testing set varies from 36.0 to 79.6 as the width of the time window goes from 1 up to 10, when absolute values are predicted from absolute values. Best results are achieved for small time windows. Similar results are achived when predicting the first difference of the signal from a window of time-lagged first differences. The best overall result of 38.4 is obtained with time window of three.

 

 

 A comparison between typical true and predicted results for linear regression on the training set is shown below. The prediction was based on a time window of size three.



The same comparison for the corresponding testing set:



• Using k nearest neighbors to predict left broca(t) from left broca(t-1,..,t-N). N varies from 1 to 10, and k varies from 1 to 7. Results show the best RMS error is achieved when predicting absolute values, and for small time windows and as many nearest neighbors as possible. The best result is 38.9, for N=2 and k=7. In contrast, predicting first differences with kNN did not perform well - the best result was 56.8, again for N=2 and k=7.

 

 
 
 
 
 

 These results can be compared against two baselines : using the mean as prediction and using the previous value in the time series. The mean value yielded a high RMSE of 49.2. When using the most recent value as a predictor, the error was somewhat lower (45.5), but not by a lot.

• More detailed results on regression This chart shows the performance of linear regression when predicting absolute values of the fifteen voxels in left Broca with a window of three time slices. Of those fifteen voxels, eight are linearly predictable, while the other seven are not.

 

 
 
 
 
 

 The regression coefficients are visualized in the image below. Black corresponds to negative and white corresponds to positive coefficients. A pixel with coordinates (x,y), counted from the upper left corner of the image, corresponds to the coefficient relating input x to output y - that is, the image corresponds to a Jacobian matrix in its layout. The time delayed vector of inputs has time slice t-N to the left and time slice t-1 to the right.



The image below shows the correlation matrix of the activation signals of the fifteen voxels in left Broca. White color corresponds to strong positive correlation, and black to no (zero) correlation.



• Using neural network to predict lb,lt(t+1) from lb,lt(t).Initial results . RMS prediction error 33.31, compared to 41.45 using means to predict.
• Using neural network to predict lt(t+1) from lb(t).Initial results . RMS prediction error 31.82, compared to 39.07 using means to predict.
• Using neural network to predict lb(t+1) from lt(t).Initial results RMS prediction error 38.89, compared to 47.1 using means to predict.

Predicting lb(t+1) from lb(t) for 02882 - PP&RRC

Method	Error on raw data	Error on percentages
linear regression	35.4030	35.3599
kernel regression	36.7323	44.2866
k nearest neighbors	37.6975	40.3757
latest value	44.3493	44.3493
mean value	47.9650	47.9650

Predicting lb(t+1) from lb(t) for 02882 - PP

Method	Error on raw data	Error on percentages
linear regression	36.9842	37.1254
kernel regression	37.1057	44.2537
k nearest neighbors	37.8392	40.1486
latest value	44.2351	44.2351
mean value	46.8361	46.8361

Predicting lb(t+1) from lb(t) for 02882 - RRC

Method	Error on raw data	Error on percentages
linear regression	37.3158	37.3919
kernel regression	38.4470	44.9196
k nearest neighbors	38.3039	40.8477
latest value	44.4042	44.4042
mean value	47.3010	47.3010

Predicting lt(t+1) from lt(t) for 02882 - PP&RRC

Method	Error on raw data	Error on percentages
linear regression	31.2933	31.1898
kernel regression	31.2634	33.4739
k nearest neighbors	32.0502	40.3757
latest value	38.5323	38.5323
mean value	39.8031	39.8031

Predicting lt(t+1) from lt(t) for 02882 - PP

Method	Error on raw data	Error on percentages
linear regression	34.2106	34.1594
kernel regression	31.8218	37.0377
k nearest neighbors	32.2334	34.2376
latest value	39.0910	39.0910
mean value	39.3279	39.3279

Predicting lt(t+1) from lt(t) for 02882 - RRC

Method	Error on raw data	Error on percentages
linear regression	34.1592	33.8042
kernel regression	31.9268	35.8551
k nearest neighbors	31.5950	33.1942
latest value	38.2995	38.2995
mean value	37.5290	37.5290

Predicting lb,lt(t+1) from lb,lt(t) for 02882 - PP&RRC

Method	Error on raw data	Error on percentages
linear regression	32.9781	32.9912
kernel regression	32.9026	38.3856
k nearest neighbors	33.4124	34.7364
latest value	40.2931	40.2931
mean value	42.3672	42.3672

Predicting lb,lt(t+1) from lb,lt(t) for 02882 - PP

Method	Error on raw data	Error on percentages
linear regression	36.6374	36.5897
kernel regression	33.5188	38.6737
k nearest neighbors	33.2764	35.0448
latest value	40.6927	40.6927
mean value	41.7719	41.7719

Predicting lb,lt(t+1) from lb,lt(t) for 02882 - RRC

Method	Error on raw data	Error on percentages
linear regression	37.4406	37.3416
kernel regression	33.8780	38.2591
k nearest neighbors	33.1496	34.6820
latest value	40.2573	40.2573
mean value	40.6477	40.6477

Predicting lb(t+1) from lb(t) for 02930 - PP&RRC

Method	Error on raw data	Error on percentages
linear regression	36.0740	36.0824
kernel regression	35.5461	39.0011
k nearest neighbors	36.2593	38.3240
latest value	44.5436	44.5436
mean value	40.2094	40.2094

Predicting lb(t+1) from lb(t) for 02930 - PP

Method	Error on raw data	Error on percentages
linear regression	37.6649	37.7794
kernel regression	36.4024	38.8822
k nearest neighbors	36.3552	38.0757
latest value	43.9044	43.9044
mean value	39.0530	39.0530

Predicting lb(t+1) from lb(t) for 02930 - RRC

Method	Error on raw data	Error on percentages
linear regression	36.4076	36.5440
kernel regression	35.7675	38.1737
k nearest neighbors	36.4065	38.1637
latest value	44.3173	44.3173
mean value	38.2857	38.2857

Predicting lt(t+1) from lt(t) for 02930 - PP&RRC

Method	Error on raw data	Error on percentages
linear regression	32.6557	32.6441
kernel regression	33.0032	36.3991
k nearest neighbors	33.9985	37.2002
latest value	38.5323	38.5323
mean value	36.9298	36.9298

Predicting lt(t+1) from lt(t) for 02930 - PP

Method	Error on raw data	Error on percentages
linear regression	33.1444	33.0677
kernel regression	33.4922	36.1160
k nearest neighbors	32.7904	35.5905
latest value	40.4429	40.4429
mean value	36.2968	36.2968

Predicting lt(t+1) from lt(t) for 02930 - RRC

Method	Error on raw data	Error on percentages
linear regression	34.1112	33.9552
kernel regression	33.1998	35.5823
k nearest neighbors	34.0388	36.9136
latest value	41.8895	41.8895
mean value	35.5599	35.5599

Predicting lb,lt(t+1) from lb,lt(t) for 02930 - PP&RRC

Method	Error on raw data	Error on percentages
linear regression	34.7935	34.8552
kernel regression	34.5333	37.8808
k nearest neighbors	34.6222	36.6398
latest value	43.4562	43.4562
mean value	38.7148	38.7148

Predicting lb,lt(t+1) from lb,lt(t) for 02930 - PP

Method	Error on raw data	Error on percentages
linear regression	37.4442	37.4967
kernel regression	35.0515	37.8006
k nearest neighbors	34.2027	36.0388
latest value	42.6260	42.6260
mean value	38.2030	38.2030

Predicting lb,lt(t+1) from lb,lt(t) for 02930 - RRC

Method	Error on raw data	Error on percentages
linear regression	36.9964	37.0235
kernel regression	34.5741	37.0835
k nearest neighbors	35.1504	37.2247
latest value	43.3474	43.3474
mean value	37.2884	37.2884

Additional information:

• Fastmap Results
• Clustering Results
• Principal Component Analysis
• Correlation / Regression Coefficients Plot
• Directory listing of /afs/cs/project/theo-35/fmri/www/

Last updated Mon Feb 1 18:13:10 EST 1999 by Daniel